List your company

Training Data Composition Estimation on RealToxicityPrompts toxic-injection study data audit (train)

22.73Estimated Toxic PercentageLLMSurgeon
7.306811.310915.31519.3191May 28, 2026

Evaluation results

MethodLinks
LLMSurgeonModel=20% Toxic GPT-2, GT Toxic (%)=20.002026.0522.7397.27
LLMSurgeonModel=10% Toxic GPT-2, GT Toxic (%)=10.002026.051298
LLMSurgeonModel=5% Toxic GPT-2, GT Toxic (%)=5.002026.057.997.1